ZONN.ai Forensic Report

Case · 86C59BF7 · IMAGE

MAnalyzed by@muzip
ZONN Analysis
0

Very likely real

Most signals point to a real, human-captured source. Detection tools are not perfect — treat this as a strong indication, not a verdict.

Signal ConfidenceLimited · 50/100

Analysed Specimen

Original analysed image
Forensic suspicion heatmap
OriginalHeatmap
POS55/100
No flagged regions

Heads up — 3 things to know

Why this analysis might be off

We highlight every disagreement and unusual signal we found so you can judge for yourself. Stronger warnings come first; informational notes are at the bottom.

AI generator fingerprints detected

AI evidence

Frequency analysis (FFT score 67/100) shows modern diffusion-style upsampling patterns, but the ML models say "real". This combination is a known blind spot for newer generators (SDXL, FLUX, Midjourney v6) — the verdict above may be misleading.

Upsampling artifacts in the frequency domain

AI evidence

FFT analysis found strong upsampling patterns — a fingerprint of diffusion-model VAE decoders (latent → pixel-space upscale).

Uneven compression (ELA)

AI evidence

ELA coefficient of variation is 1.52 — different regions show noticeably different compression levels. Common with composites, edits, or AI generations.

Origin Check

Trace this image elsewhere

Cross-reference the source against major reverse-image services. Each link opens in a new tab with the image URL preloaded — ZONN.ai does not re-upload the image.

Why this verdict

  • CommFor (4803 Generators)read real · 0/100

    CommFor detector trained across 4,803 generator variants for broad coverage.

  • Error Level Analysisread real · 0/100

    Error Level Analysis. Re-saves and diffs to expose uneven compression regions.

Model Agreement

42%

Variance across 6 ML detectors. Higher agreement means the models converged on the same reading; lower agreement means treat the verdict with care.

Evidence — 16 detectors reviewed

What each detector saw

Each detector independently gave this imagea score from 0 (definitely real) to 100 (definitely AI). The score above is their weighted consensus — detectors with higher confidence count more. No single detector decides; you read the spread.

ML Models6 detectors · mean 24
▸ expand
CommFor (4803 Generators)
0
INA v2 (FLUX/MJ)
1
xRayon ConvNeXtV2
3
SigLIP AI Detector
12
Bombek1 SigLIP+DINOv2
75
Manipulation Map (IML-ViT)
50
Pixel & Frequency Forensics7 detectors · mean 46
▸ expand
Error Level Analysis
0
Pixel Analysis
75
Noise Pattern
68
Frequency Analysis
67
Edge Consistency
35
Color Distribution
40
Compression Quality
40
Provenance & Metadata3 detectors · mean 54
▸ expand
ICC Profile
62
Metadata
50
C2PA Provenance
50

Image Quality

Dimensions
800 × 600 px
Aspect
1.333
File size
34.8 KB
Bytes / pixel
0.074

Frequency Analysis

Radial1.000
DCT0.676
Upsampling1.000
Cross-channel0.009
Power-law β
-3.21
Grid energy
0.487

Edge Consistency

CV 1.167
Cell 1: 0.3117Cell 2: 0.2806Cell 3: 0.2222Cell 4: 0.2126Cell 5: 0.5928Cell 6: 0.7639Cell 7: 0.8510Cell 8: 0.7027Cell 9: 11.0114Cell 10: 10.7573Cell 11: 18.9001Cell 12: 9.2325Cell 13: 34.0906Cell 14: 34.2130Cell 15: 34.0400Cell 16: 34.7020

Per-region edge density (4 × 4 grid). Uneven distribution may indicate localized editing or splicing; uniform fields are typical of fully synthetic outputs.

Range: 0.212634.7020

Noise Fingerprint

Variance
78.49
Std deviation
8.86
Mean
-0.0
Spatial corr.
2.944
Mean Δ
0.68
σ
1.03
CV
1.519
Uniformity
-0.519

Provenance

Source Dossier

PlatformDirect upload
Author
Content Typeimage
Analyzed OnMay 18, 2026, 5:14 PM
Analyzed by@muzip